An attention-assisted detection method for anomalous events in social networks

The widespread use of online social networks is accompanied by frequent abnormal events that spread rapidly. This poses a threat to online platforms and the public opinion environment. The persistent evolution and bursty spread of abnormal events on social networks make it difficult to capture the structural and temporal patterns of this ever-changing process. Existing approaches face three challenges: deep semantic mining in evolving events, dynamic modeling in detection, and the explainability for social network event detection. To address these, this paper proposes an attention-assisted detection method for abnormal events in social networks. Based on semantics, attention, and temporal information of user-generated text, the proposed methodology constructs a social network temporal graph model. This dynamic graph is better suited to discovering abnormal patterns by considering possible hints, including structural, content, and temporal features. An algorithm for social network event attention degree is then designed to detect bursts of social network events. Next, the method combines the graph attention network and density clustering to perceive the depth semantic themes of events. This process reveals hidden anomalous topics and constructs scenarios as events evolve, thereby improving the interpretability of anomalous events in social networks. Finally, the dynamic evolution stages of abnormal events are analyzed based on the velocity and acceleration of attention degree. Validation experiments are conducted on public datasets, and the case-based test results show that the core evaluation indicators of the proposed method for detecting abnormal events exceed 0.92.